Moonshot AI model guide

Kimi K2.6

Kimi K2.6 is Moonshot AI's multimodal K2 model for agentic coding, long-context reasoning, visual understanding, and long-running agent tasks. Official Kimi API docs list a 262,144-token context window, text/image/video input, thinking and non-thinking modes, and CNY pricing per 1M tokens.

262,144
context tokens
58.6
SWE-Bench Pro
CNY 27
output / 1M tokens
Capabilities

What Kimi K2.6 Is Built to Do

The official materials position Kimi K2.6 around agentic coding, multimodal understanding, long thinking, and longer autonomous execution rather than short single-turn chat alone.

Long-horizon coding

Moonshot reports Kimi K2.6 running a 12+ hour local inference optimization with 4,000+ tool calls and 14 iterations, improving throughput from about 15 to 193 tokens per second. In another 13-hour exchange-core task, it used 1,000+ tool calls, modified 4,000+ lines, and improved throughput by 185% and 133% on two reported measures.

Multimodal tool loops

The Kimi API docs show image and video input examples, multimodal tool results, ToolCalls, JSON Mode, Partial Mode, automatic context caching, official web search, and a thinking switch for K2.6.

Agent swarm and background work

The K2.6 blog describes agent-swarm coordination up to 300 sub-agents and 4,000 coordinated steps, plus a 5-day autonomous engineering worklog for monitoring, incident response, and system operations.

Kimi K2.6 benchmarks

Kimi K2.6 Benchmarks from the Official Technical Blog

These rows use the public Kimi K2.6 benchmark table. They are useful selection signals for coding, agentic search, long-context tool use, visual reasoning, and general reasoning, but they should still be validated on your own prompts and tools.

Kimi K2.6 benchmarksCategoryKimi K2.6 scoreWhat it indicates
HLE-Full w/ toolsAgentic reasoning54.0Full Humanity's Last Exam with tools in the official Kimi table.
BrowseCompAgentic search83.2Autonomous browsing and hard web-retrieval capability.
DeepSearchQADeep retrieval92.5 F1 / 83.0 accuracyAgent deep-search answer quality in Moonshot's table.
Terminal-Bench 2.0 (Terminus-2)Coding agents66.7Terminal-based task execution under the Terminus-2 harness.
SWE-Bench ProSoftware engineering58.6Realistic software engineering task resolution signal.
SWE-Bench VerifiedSoftware engineering80.2Verified software issue resolution benchmark result.
LiveCodeBench v6Coding89.6Competitive programming and code-generation signal.
MMMU-ProVision79.4Multimodal university-level reasoning benchmark.
MathVision w/ pythonVisual reasoning93.2Visual math reasoning with Python tool use.
V* w/ pythonVisual reasoning96.9Visual reasoning task with Python tool support.

Moonshot states that, unless otherwise specified, Kimi K2.6 experiments used temperature = 1.0, top-p = 1.0, and a 262,144-token context length. It also says coding scores are averaged over 10 independent runs and vision benchmarks over three runs.

Comparison point

Kimi K2.6 vs GLM-5.1, Opus 4.7, and DeepSeek V4 Pro

The comparison below keeps vendor-reported data separated by source. It does not pretend that every number comes from one identical harness; where a vendor publishes only specs or internal evals, the table says so.

Comparison pointKimi K2.6GLM-5.1Claude Opus 4.7DeepSeek V4 Pro
Best fitMultimodal coding agents, visual work, deep search, and long-context tool loops.Text-only long-horizon engineering agents and sustained coding execution.Closed flagship for difficult software engineering and high-resolution image work.Open-weight million-token reasoning, coding, and high-volume API use.
Benchmark signalsSWE-Pro 58.6, Terminal 66.7, BrowseComp 83.2, DeepSearchQA F1 92.5.SWE-Pro 58.4, Terminal 63.5, BrowseComp 79.3, MCP-Atlas 71.8.Anthropic reports +13% vs Opus 4.6 on a 93-task coding eval.LiveCodeBench 93.5, Terminal 67.9, SWE Verified 80.6, BrowseComp 83.4.
Context and output262,144 context; 32,768 default max_tokens in Kimi docs.200K context; 128K maximum output.1M context at standard Opus pricing; verify output limits.1M context; 384K maximum output.
Pricing postureCNY 1.10 cached input, CNY 6.50 input, CNY 27 output.$0.26 cached input, $1.40 input, $4.40 output.$5 input, $25 output; separate cache rates.$0.003625 cached input, $0.435 input, $0.87 output.
Modality and toolsText, image, video, thinking mode, tools, JSON, cache, and web search.Text, thinking mode, function calling, cache, structured output, and MCP.Text/image input, higher-resolution vision, effort control, and Claude tools.Text, thinking modes, JSON, tool calls, chat prefix, and FIM beta.
Model size and accessK2 lineage: 1T total / 32B active; K2.6 count not separately disclosed.754B parameters; MIT license.Closed model; parameter count not published.1.6T total / 49B active; MIT license.
Kimi K2.6 pricing

Official Kimi K2.6 API Pricing

Kimi prices K2.6 in CNY per 1M tokens. The official table separates cache-hit input, cache-miss input, and output tokens, and lists the context window as 262,144 tokens.

Cache-hit input
CNY 1.10 / 1M tokens
Cache-miss input
CNY 6.50 / 1M tokens
Output
CNY 27.00 / 1M tokens
Context window
262,144 tokens

Prices are taken from the official Kimi K2.6 pricing page checked on June 25, 2026. Currency conversion, MyClaw plan packaging, provider routing, and promotions can change the effective production cost.

MyClaw workflows

Where Kimi K2.6 Fits inside MyClaw Agents

Choose Kimi K2.6 when the agent needs multimodal evidence, long reasoning traces, coding tools, and enough context to keep a complicated task coherent.

Repository implementation

Use Kimi K2.6 for codebase reading, issue reproduction, file edits, terminal feedback, and long-running repair loops where the model must keep its plan stable.

Visual agent tasks

Bring screenshots, images, and video clips into the same agent loop as text prompts and tool outputs, then ask Kimi to reason over what it sees.

Deep search and research

The official DeepSearchQA, BrowseComp, and HLE-with-tools results make K2.6 a strong candidate for search-heavy research agents that need source gathering plus synthesis.

Model comparison and routing

Compare Kimi K2.6 against GLM-5.1, Opus 4.7, and DeepSeek V4 Pro on the same MyClaw task, then route production work by modality, context, cost, and reliability.

More models

Explore Other AI Models in MyClaw

Compare Kimi K2.6 with other model pages for coding, reasoning, context length, multimodal work, and production routing.

Kimi K2.6 FAQs

Evaluate Kimi K2.6 in a Hosted MyClaw Agent

Use MyClaw to test Kimi K2.6 on repository work, visual evidence, web research, long-context prompts, and tool-heavy agent loops before choosing your production model mix.

Run Kimi K2.6 in MyClaw